Unsupervised Graph Embeddings for Session-based Recommendation with Item Features

Andreas Peintner, Marta Moscati, Emilia Parada-Cabaleiro, Markus Schedl, Eva ZangerlearXiv 2025 · DOI · arXiv

TL;DR

This paper introduces GCNext, a novel method that enhances session-based recommendation systems by leveraging item features within a graph structure. By creating a rich item co-occurrence graph and learning unsupervised item embeddings, GCNext can be easily integrated into existing recommendation models, significantly boosting their performance.

What Problem Does This Paper Address?

Session-based recommendation systems aim to predict a user's next action (e.g., clicking an item) based solely on their recent interactions within a current session. This is particularly challenging because user identities are often unknown, and historical data beyond the current session is unavailable. While existing methods use Recurrent Neural Networks (RNNs), Graph Neural Networks (GNNs), or attention mechanisms to model sequential patterns, they often struggle to effectively incorporate auxiliary item features. This paper addresses the gap of seamlessly integrating rich item features into session-based recommendation models, especially within graph-based approaches, to improve prediction accuracy.

Methods

The core of the proposed method, GCNext, involves two main phases:

  1. Unsupervised Graph Embeddings:
    • An item co-occurrence graph is constructed where nodes represent items and edges represent their co-occurrence within sessions. Edge weights are normalized co-occurrence counts.
  2. Item features (e.g., category, price) are attached to each node.
  3. Graph Convolutional Networks (GCNs), specifically using a GATv2 variant with attentional convolution, are employed to learn unsupervised node (item) embeddings from this feature-rich graph.
  4. To handle large graphs that may not fit into GPU memory, a sub-graph sampling approach based on neighborhood batch sampling is used.
  5. The learning method is based on Bootstrapped Graph Latents (BGRL), which uses two encoders to generate online and target embeddings and optimizes for cosine similarity between a predictor's output and the target embedding. This approach avoids the need for explicit negative sampling, making it more scalable.
  6. Integration with Sequential Models:
    • For Neural Network Models: The learned item embeddings from GCNext are used to initialize the item embedding table of existing sequential models (e.g., GRU4Rec+, Caser, STAMP, SASRec). This provides a richer starting point for the model's learning process.
  7. For Nearest-Neighbor Methods: GCNext is used to compute the similarity between sessions based on the cosine distance of item embeddings. This allows nearest-neighbor methods to find more relevant similar sessions.

The key novelty lies in extracting item embeddings from a feature-rich graph before feeding them into sequential models, rather than trying to integrate features directly within the sequential model's architecture. This makes GCNext a flexible, plug-in extension.

Datasets

The authors evaluated GCNext on three publicly available datasets:

  • Diginetica: An e-commerce dataset with item category and price as features.
  • Tmall: An e-commerce dataset with category, brand, and seller as item features.
  • Music4All+: A music dataset enriched with i-vectors derived from Mel-Frequency Cepstral Coefficients, providing 11 item features.

The datasets were preprocessed by filtering out infrequent items and short sessions, and sequences were split into smaller subsequences for training.

Key Insights and Ideas

  • Item features are crucial for session-based recommendations: The paper strongly suggests that incorporating item features, especially through a graph structure, significantly enhances performance beyond just modeling sequential item transitions.
  • Unsupervised graph embeddings offer a powerful, flexible enhancement: GCNext demonstrates that learning item representations from a co-occurrence graph with features in an unsupervised manner provides a valuable boost to various existing recommendation models without requiring architectural changes.
  • Plug-in nature is a major advantage: The ability to easily integrate GCNext into both nearest-neighbor and neural network models without modifying their core architectures makes it highly practical for real-world applications.
  • Graph structure captures valuable item relationships: The item co-occurrence graph, enriched with features, seems to capture latent relationships between items that are beneficial for predicting user intent.
  • Faster convergence for neural networks: Models initialized with GCNext embeddings converge faster, likely due to the richer, pre-learned information.

Key Results

  • GCNext consistently improved the performance of both nearest-neighbor and neural network-based session recommendation models across all three datasets.
  • Significant improvements were observed: For instance, STAMP extended with GCNext achieved a performance increase of up to 12.79% on the MRR@20 metric.
  • On the Diginetica dataset, GCNext combined with SASRec (a transformer-based model) achieved the highest scores across all metrics.
  • For the Tmall dataset, SASRec with GCNext improved the state-of-the-art HR@10 score by 1.12%.
  • Nearest-neighbor methods, when enhanced by GCNext, showed performance gains, with V-SKNN improving by 0.88% to 6.33% on the HR@10 score.
  • The paper also notes that neural network models trained with GCNext embeddings converged faster than their standard counterparts.

Limitations and Future Work

  • Cold-start scenarios: The authors plan to further investigate the potential improvements of graph-based item embeddings in cold-start scenarios for session-based recommendation.
  • Comparison with other graph embedding methods: Future work will involve comparing GCNext with different graph embedding methods.
  • Diverging impact on baseline models: Investigating how GCNext impacts different baseline models differently is also a planned future direction.
  • Scalability for extremely large graphs: While sub-graph sampling is used, the paper implicitly acknowledges that handling graphs with hundreds of millions of nodes can still be a challenge, though BGRL helps mitigate this.

Summary updated

Abstract

In session-based recommender systems, predictions are based on the user's preceding behavior in the session. State-of-the-art sequential recommendation algorithms either use graph neural networks to model sessions in a graph or leverage the similarity of sessions by exploiting item features. In this paper, we combine these two approaches and propose a novel method, Graph Convolutional Network Extension (GCNext), which incorporates item features directly into the graph representation via graph convolutional networks. GCNext creates a feature-rich item co-occurrence graph and learns the corresponding item embeddings in an unsupervised manner. We show on three datasets that integrating GCNext into sequential recommendation algorithms significantly boosts the performance of nearest-neighbor methods as well as neural network models. Our flexible extension is easy to incorporate in state-of-the-art methods and increases the MRR@20 by up to 12.79%.
BibTeX
@misc{peintner2025unsupervised,
  author = {Andreas Peintner and Marta Moscati and Emilia Parada-Cabaleiro and Markus Schedl and Eva Zangerle},
  title = {{Unsupervised Graph Embeddings for Session-based Recommendation with Item Features}},
  year = {2025},
  eprint = {2502.13763},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2502.13763},
}
Figure